Core Concepts
CFEAR is an efficient and accurate method for spinning 2D radar odometry that generalizes well across environments. This article presents an evaluation of CFEAR on the public Boreas dataset, demonstrating low drift and real-time performance.
Abstract
The article describes the CFEAR Radar Odometry method, which was submitted to a competition at the Radar in Robotics workshop at ICRA 2024. CFEAR is designed to be efficient and accurate for spinning 2D radar odometry, generalizing well across diverse environments.
The key highlights and insights from the article are:
CFEAR uses a two-step feature extraction approach, first filtering the radar data conservatively and then extracting a sparse but descriptive representation of the scene.
The odometry estimation is performed by minimizing the distance between corresponding surface points, with outlier rejection using a robust loss function and a similarity heuristic.
A significant change to the previous implementation is the replacement of the k-d tree with a hash table for rapid neighboring surface point lookup, enabling the deployment of CFEAR with significantly more keyframes in real-time.
A coarse-to-fine strategy is employed to reduce the radius of association and increase outlier rejection in later iterations, which helps address rare failures during rapid turns.
Experiments on the Boreas dataset show that the CFEAR-CTF-S10 configuration, with 10 keyframes, reaches as low as 0.66% translation drift at a frame rate of 68 Hz.
Additional experiments on the Oxford and MulRan datasets demonstrate the high level of generalization of CFEAR, with 1.16% and 1.18% drift respectively, without any parameter tuning.
The authors note that future work should investigate the cause of systematic trajectory errors observed in the experiments.
Stats
The article presents the following key metrics:
0.66% translation drift at 68 Hz for the CFEAR-CTF-S10 configuration on the Boreas dataset training set.
1.16% translation drift on the Oxford dataset sequences.
1.18% translation drift on the MulRan dataset sequences.
Quotes
"Surprisingly CFEAR-CTF-S10 reached as low as 0.66% in the Boreas training set."
"Larger improvements were observed when experimenting with other parameter sets. The major improvement is attributed to the use of additional keyframes which is possible with the significantly more efficient nearest neighbor search."